Analysis updated 2026-05-18
Reproduce a research model that forecasts macroeconomic indicators from real and synthetic data.
Pretrain a small forecasting backbone on synthetic time series data before fine-tuning it.
Run zero-shot forecasts on economic time series using the trained MacroCast checkpoints.
| shubhranshu-shekhar/macrocast | 0xhassaan/nn-from-scratch | 3ks/embedoc | |
|---|---|---|---|
| Stars | 0 | 0 | — |
| Language | Python | Python | Python |
| Last pushed | — | — | 2023-06-08 |
| Maintenance | — | — | Dormant |
| Setup difficulty | hard | moderate | hard |
| Complexity | 5/5 | 4/5 | 1/5 |
| Audience | researcher | developer | developer |
Figures from each repo's GitHub metadata at analysis time.
Requires a CUDA-capable GPU, PyTorch, and access to FRED-MD vintage data to reproduce the fine-tuning step.
MacroCast is code for reproducing a machine learning model built to forecast macroeconomic time series, meaning things like inflation, employment, or output that economists track over time. The project explains the model as coming together in two steps. The first step, called Forge, trains a small neural network, about 1.2 million parameters in size, using only made up, synthetic data rather than any real economic numbers. This backbone is built on top of an existing published model called TempoPFN. The second step takes that pretrained Forge model and fine-tunes it further using real historical economic data from a dataset called FRED-MD, combined with additional synthetic data that mimics realistic economic patterns. The end result of this second step is the actual MacroCast model. The README describes the model's structure at a glance: it uses a type of recurrent neural network layer, produces forecasts as nine different quantiles rather than a single number, meaning it gives a range of plausible outcomes instead of one guess, and it forecasts a full sequence at once rather than working in small overlapping windows. The synthetic data used to pretrain Forge comes from several different mathematical processes, while the mixture used to fine-tune MacroCast blends real data with several statistical modeling techniques used in economics, such as autoregressive models and factor models. Documentation for each of the two steps, along with a separate guide for running forecasts once the model is trained, lives in their own folders within the repository. The project is built directly on top of TempoPFN's code, including its model backbone, its training system, and its synthetic data generators, with the borrowed code kept in its own folder and credited under its original Apache 2.0 license. The authors also credit a separate project called Flash Linear Attention as the basis for one of the model's layers. MacroCast itself is released under the Apache License 2.0, matching the license of the code it depends on. The README includes citation information for anyone who wants to reference this project or the underlying TempoPFN model in academic work.
Research code for a two-step neural network model that forecasts economic indicators like inflation and employment, trained on synthetic and real data.
Mainly Python. The stack also includes Python, PyTorch.
Use freely for any purpose, including commercial use, as long as you keep the copyright notice.
Setup difficulty is rated hard, with roughly 1day+ to a first successful run.
Mainly researcher.
This repo across BitVibe Labs
Verify against the repo before relying on details.